The world of fundamental mathematics stands on the brink of a tectonic shift. The internal build of OpenAI's next-generation AI model Astra has demonstrated results previously considered unattainable for machine intelligence: 10 new solutions or significant progress on long-standing open problems in high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. There is currently no public access to this model, which only fuels interest in the announcement.
What exactly did OpenAI publish
The company presented a 249-page technical report, detailed breakdowns of reasoning processes, and a repository with proof formalizations in Lean 4—a formal verification system where every logical step is machine-verified. According to the developers, the mathematical arguments were generated specifically by the internal version of Astra. Humans were involved only in preparing manuscripts and collaborative refinement, after which the model independently formalized each argument into a Lean certificate.
"We believe that attribution should honestly reflect how the result was obtained: attributing authorship of a proof fully generated by an AI system to a human would distort the system's contribution and the nature of genuine human intellectual work," OpenAI stated.
The economic side of the issue is also intriguing. OpenAI claims that the total volume of tokens needed to find solutions would cost about $2000 at Sol API rates. However, this estimate does not account for subsequent verification, formatting, and the work of the mathematical community. The company also remains silent about the number of failed runs or unsolved problems left out of the publication.
Quantum breakthrough and post-quantum security
One of the key results concerns quantum parallel repetition—a characteristic of how the probability of success changes when the same interactive game is run multiple times simultaneously. In classical theory, repetition sharply reduces the chance of cheating, but in the quantum case, the problem becomes more complex due to entanglement. Astra proved exponential parallel repetition for arbitrary finite two-player quantum games. If the result passes independent verification, it could have serious implications for quantum verification, cryptographic protocols, and interactive proofs.
Another significant result relates to the closest vector problem—finding the nearest lattice point to a given point in space. This is the foundation of lattice cryptography, one of the main directions of post-quantum protection. According to OpenAI, Astra obtained a stronger result on the complexity of approximately solving this problem. It is important to emphasize: this does not mean an attack on already deployed post-quantum systems, but it opens new horizons for decoding and lattice-related norms.
Other achievements and their significance
OpenAI calls the construction of a non-sofic group—an object whose existence question had remained open since the late 1990s—the main result of the publication. Among other claimed achievements:
- improved estimates for sphere packing in high dimensions;
- stronger bounds for binary and spherical codes;
- new lower bounds in arithmetic circuit complexity;
- a result on Ehrhart's volume conjecture;
- a counterexample to Connes's rigidity conjecture;
- new results on multicolor Ramsey numbers;
- counterexamples to two Erdős conjectures in extremal graph theory.
These achievements belong to fundamental mathematics, and their practical impact is not immediate. However, the publication itself intensifies the discussion about whether AI models can not only assist researchers but also independently produce new scientific results.
Why verification is needed
Formalization in Lean reduces the risk of ordinary logical gaps: the published code can be compiled and machine-verified. However, it does not replace scientific expertise. Mathematicians will need to assess how accurately the formal statements correspond to the original open problems, how general the obtained results are, and what place they occupy in the existing literature. This context is especially important after the publication of the Leiden Declaration on AI and Mathematics, endorsed by the International Mathematical Union. The document calls for preserving peer review, expert scrutiny, and transparent attribution when using artificial intelligence in mathematical research.
OpenAI directly referenced the declaration and stated that it takes responsibility for the correctness of the published materials but attributes the mathematical arguments themselves to the system, not to the people who helped format the manuscripts. However, the company did not announce a public launch date for Astra and did not provide the model for independent testing. The reproducibility of the claimed process cannot yet be verified.
Recall that in May, an OpenAI model already disproved the 80-year-old Erdős conjecture on unit distances, and a few months later, Anthropic researcher Levent Alpoge used Claude Fable 5 to find a counterexample to the Jacobian conjecture. These events are shaping a new trend that cannot be ignored.
My view: OpenAI's announcement is not just another milestone in AI development, but a signal that we are approaching a point where machines become full participants in the scientific process. However, the market and the academic community should maintain healthy skepticism: until the results are reproduced and verified by independent experts, their status remains preliminary. Nevertheless, if these proofs are confirmed, we will witness the beginning of a new era in mathematics—and possibly in cryptography.